A fallen power line in Virginia knocked more than 3 gigawatts of data-center load offline in under 30 seconds. It will not be the last time.
On July 22, 2026, a single transmission line fault outside Washington, D.C. set off a chain reaction that grid engineers had been warning about for two years. Data centers across Northern Virginia's "Data Center Alley," the densest concentration of server capacity on the planet, sensed the disturbance and switched to backup power almost simultaneously. More than 3.1 gigawatts of load, roughly 3% of the entire PJM grid's demand at that moment, dropped off within about 30 seconds. The grid took ten to eleven minutes to restabilize, far longer than the millisecond-scale corrections operators are used to.
Nobody lost power in a sustained way. But lights flickered from Virginia to Chicago, and the event was roughly double the size of a similar disconnection in 2024. Grid-reliability filings and data-center power design analysis over the past several years point to this incident as a turning point in how seriously utilities, regulators, and the industry itself are treating AI's electrical footprint. This is no longer a theoretical risk. It is a recurring operational event with a growing paper trail.
Quick Answer
AI data center power outages are less often full blackouts and more often rapid disconnections to backup power, triggered when a grid disturbance causes protective systems to isolate a facility. These events have grown in scale, from about 1.5 GW in 2024 to over 3 GW in July 2026, and grid operators including PJM and ERCOT are now writing new rules, including possible data-center curtailment starting mid-2027, to keep them from cascading into broader failures.
Here is what actually happened, why AI workloads are different from traditional data-center load, and what regulators and grid operators are doing about it.
Inside the July 2026 Virginia Disconnection
The sequence started with something mundane: a downed power line outside Washington, D.C. In a conventional grid, a single line fault causes a brief voltage dip that most equipment rides through without issue. Data centers running AI workloads are not built the same way. Many are configured to protect sensitive GPU clusters by disconnecting from utility power and switching to on-site backup the moment voltage or frequency drifts outside a narrow band.
That protective instinct, multiplied across dozens of facilities in the same few square miles, is what turned a routine fault into a grid event. Dominion Energy, the utility serving the region, said its own systems did not shed the load. The data centers disconnected themselves. The result was a temporary generation surplus of nearly 3.5 GW, a voltage spike that traveled hundreds of miles, and a stabilization period nearly a hundred times longer than a typical correction.
Scale of Recent Data-Center Disconnection Events
Each disconnection has been larger than the one before it. NERC's 2026 State of Reliability report lists these among the near-term risks it is tracking most closely.
Why AI Workloads Are Harder on the Grid
Traditional data centers draw a relatively flat, predictable load. Training and inference workloads for large AI models do not behave that way. Power draw can spike as much as 50% above a facility's design capacity within seconds as thousands of GPUs shift between idle and full utilization in near-unison. Bloomberg's reporting on this pattern in August 2026 described batteries, generators, and cooling systems being pushed well outside the operating envelope they were designed for, leading to accelerated wear and, in some cases, outright malfunction.
The financial exposure compounds the technical one. Multibillion-dollar facilities lose meaningful revenue during even brief downtime, and investors are increasingly asking how much faster this volatility is aging equipment that was underwritten on a longer depreciation schedule. A quick way to see how sensitive a project's return actually is to that kind of downtime and rate exposure is running the numbers through a debt coverage ratio calculator before financing is finalized. Wood Mackenzie's mid-2026 commentary noted that two of a still-small number of behind-the-meter AI data-center sites have already suffered serious incidents, in one case turbine damage running into the millions of dollars, and in another a complete site blackout.
| Equipment | Stress From AI Load Swings |
|---|---|
| On-site generators | Rapid start-stop cycling and sustained partial loading accelerate mechanical wear |
| Battery / UPS systems | Frequent deep charge-discharge cycles shorten usable battery life |
| Cooling systems | Sudden thermal load swings strain compressors and control systems |
| Grid transformers | Voltage and frequency disturbances can propagate to nearby residential equipment |
For operators trying to size backup capacity or cooling against a realistic load profile rather than a nameplate rating, a data center power calculator is a useful starting point, and a PUE calculator helps benchmark how much of that power draw is actually reaching compute versus cooling and overhead.
How Regulators and Grid Operators Are Responding
Grid Reliability Risk From AI Load Swings
Operators describe current events as manageable disturbances, not system failures. NERC and multiple grid analysts nonetheless place near-term reliability risk in the moderate range, rising if disconnection events keep growing in scale before ride-through rules take full effect.
The Slower-Moving Problem Behind the Headlines
Disconnection events get the attention, but the deeper structural issue is how long it now takes to get a new data center connected to the grid at all. The national interconnection queue has swelled to roughly 2,600 gigawatts of proposed generation and storage, more than the country's entire existing operational capacity, and the median wait time from request to commercial operation is approaching five years nationally.
Grid Interconnection Wait Times by Market (2026)
A campus that enters the queue today in one of the top data-center markets generally cannot expect utility power for four to seven years, regardless of how quickly the building itself is finished.
Large power transformers are a big part of why: lead times for this equipment have climbed from roughly 140 weeks in 2023 to more than 160 weeks in 2026, and nearly 80% of projects that eventually withdraw from interconnection queues do so because of unpredictable, multi-year delays and grid-upgrade costs that can run 30 to 37% of a project's total budget. This is reshaping where new capacity gets built and why. Markets with existing headroom, cooperative utilities, and shorter queues are increasingly winning projects that would once have defaulted to Northern Virginia by habit. A closer look at data center site selection criteria in the United States and a deeper breakdown of grid connection problems facing new data centers covers how developers are adjusting.
Who Feels the Impact
Utilities and grid operators
Face the operational burden of stabilizing sudden multi-gigawatt swings and the political burden of capacity auctions that are already coming up short of projected demand.
Data center operators
Larger hyperscalers can absorb the cost of dedicated backup generation and on-site buffering. Smaller, newer facilities without contracted power are the ones most exposed to future curtailment rules.
Residential customers
Generally do not lose power outright, but can see flickering lights, appliance noise from voltage disturbances, and rising bills as capacity costs get passed through.
The clearest emerging solution is not a single fix but a combination: better ride-through capability so facilities do not disconnect at the first disturbance, more on-site generation and battery buffering so AI-driven swings never reach the grid in the first place, improved modeling so operators can anticipate cascading risk before it happens, and cost allocation frameworks that ask large loads to help pay for the capacity their growth requires.
Bottom Line
No AI data center outage in 2026 has caused a prolonged, widespread blackout. What has happened is a series of increasingly large, increasingly frequent disconnections, each one a warning sign that AI's power demand is outrunning the grid's ability to absorb it smoothly. The July 2026 Virginia event was twice the scale of 2024's. If that trend continues while ride-through rules and dedicated generation commitments are still being phased in, the industry's own "canary in the coal mine" framing looks less like caution and more like an accurate description of where this is headed.
Related Reads
This article reflects reporting and regulatory filings published through August 2026 from sources including TechCrunch, Reuters, Bloomberg, Dominion Energy, PJM Interconnection, NERC, Data Center Knowledge, Latitude Media, and Wood Mackenzie. Grid conditions and regulatory rules are evolving; check the latest PJM, ERCOT, and NERC updates for current status.
Core Insights Review contributors publish research-based analysis and editorial insights on commercial real estate, PropTech, smart infrastructure, sustainable construction, industrial real estate, and emerging technologies shaping the future of the built environment.
Check for more information: Core Insights Review
